Knowledge extraction. WP4 WP4 aims at exploiting the development and results of WP3 and WP5 to provide techniques for the semantic analysis, annotation and retrieval of multimedia content. The overall approach will be based on knowledge assisted content analysis and annotation using a multimedia ontology infrastructure. More specifically, semantic and low-level attributes of the objects to be detected in combination with appropriately defined rules developed within WP5 will determine the set of algorithms, which will be the aim of WP4, and parameters required for the detection of semantic objects. Semantic concepts within the context of specific domains will be defined in ontologies, extended with qualitative attributes of the semantic objects (e.g. colour homogeneity), multimedia processing methods (e.g. colour clustering), and numerical data or low- level features (e.g. colour models, also defined in the ontology). Semantic Web technologies will be used for knowledge representation including rules to describe how tools for multimedia analysis should be applied according to different object attributes and low-level features as extracted in WP3. A detailed description of the research plan for the first phase of the project (months 1- 18) is given in section
Knowledge extraction. Historically, different approaches have been used to extract information from a data set. In the simplest cases, names, time, space, and classes can be ignored, and extracting knowledge from a data set I is, therefore, reduced to understanding the mathematical behavior of the attributes, often referred to as descriptive statistics. In doing so, we analyze each attribute (or variable) independently, and the goal is to summarize the information that we can extract from such an analysis. We use descriptive statistics to describe individual quantitative observations and, very often, the methods that are applied are graphical by nature (i.e. histograms, boxplots, and scatter plots). Descriptive statistics are progressively integrated with inferential statistics when relations between variables are taken into account, and even in simple scenarios, much effort can be devoted to understanding the possible multivariate relationships that the data present. As indicated previously, such relationships may help to understand how the data behave. When considering classes, the idea of causality arises, and regression and classification models are used to understand the processes that underlie the data. Regression is the term commonly used to denote the inferential methods for describing the behavior of one variable (for this study, the class variable when it is numeric) in terms of other variables (for this study, the attribute variables). Classification, on the other hand, deals with non-numerical classes, and consists of describing how the class variables depend on the values of the attributes. These two problems originally stem from different areas (inferential statistics the former, machine learning the latter) but share a number of concepts and methods. It can be argued that, as of late, machine learning when intended as a discipline, includes both regression and classification tasks and algorithms. Moreover, the ever-growing quantity of data that are produced and recorded, and that are therefore object of analysis, justifies a frequentist approach to learning – typically associated with machine learning, as opposed to a distribution-based – typically associated to inferential statistics. Interestingly, focusing on classification does not automatically entail that neither descriptive or simple inferential methods become ineffectual. For example, a simple binary classification problem can be approached by observing how the mean of a particular attribute changes amo...
